[Lynxifyy]

What changed

The preloader had no CI at all, so nothing checked that the tree still builds, that the host
suite still passes, or that the sources listed in CMakeLists.txt are the ones on disk. This
adds four jobs, each a real check rather than a smoke test. A green run is the claim that the
library a client loads is the library the sources describe.

Job	What it proves
host-tests	The 57-test C++ suite passes, including the three ThreadSanitizer tests.
rust	cargo fmt --check, clippy with the correctness group as errors, and the 249-test workspace suite.
native-build	libsweetloader.so and the three Rust cdylibs cross-compile for arm64-v8a, and the built .so exports the JNI and SDK symbols with no static-archive leakage.
repo-integrity	No .cpp missing from SWEETLOADER_SOURCES, no host test unregistered, every fetched dependency pinned to a commit and attributed in NOTICE.

The first run failed, and it was worth having

The first run went red on three of the four jobs. Every failure was a defect in the gate itself,
which is the failure mode a gate exists to prevent, so each is worth naming:

native-build claimed a present symbol was missing. The check was
llvm-nm -D --defined-only libsweetloader.so | grep -q " GetSweetLoaderInput$". Under
set -o pipefail, grep -q exits on its first match and closes the pipe, the producer takes
SIGPIPE, and the pipeline reports that instead of the match. GetSweetLoaderInput is in the
dynamic table; the check said it was not. Reproduced directly:

$ ( set -e;   nm -D --defined-only libsweetloader.so | grep -q " GetSweetLoaderInput$" ) ; echo $?
0
$ ( set -euo pipefail; nm -D --defined-only libsweetloader.so | grep -q " GetSweetLoaderInput$" ) ; echo $?
74          # 128 + SIGPIPE

The symbol tables are now written to files once and every check greps a file.

host-tests could not create its cache. run_tests.sh cached its fetched nlohmann/json
and fmt headers four levels above its own root:

mkdir: cannot create directory '/home/runner/work/SweetLoader/SweetLoader/../../../../build': Permission denied

That path is correct for this tree as the client's submodule and escapes the checkout entirely
for a standalone clone, which is what CI has. The cache now sits beside the checkout and honours
SWEETLOADER_HOST_CACHE, which the job points at the runner's temp dir.

rust was right: the tree was not formatted. cargo fmt --check reported 157 sites across
30 files, because the Rust tree had never been through rustfmt. No configuration rescues it
(max_width 100 or 120, with and without use_small_heuristics = "Max", all leave more than 120
sites), so the tree is normalized in this PR, with the test suite and clippy re-run green after.

A fourth fix was needed after the tree was normalized. SWEETLOADER_HOST_CACHE was first
placed in the job's env:, where the runner context is not in scope. GitHub rejected the whole
file, so the run reported failure with zero jobs and no annotation. actionlint names it:
context "runner" is not allowed here. The variable moved to the step, where it is in scope.

Why each remaining gate is shaped the way it is

The sanitizer tests must actually run. run_tests.sh skips a ThreadSanitizer test when
setarch is unavailable, and a skip reads as a pass. The runner has ASLR on and TSAN refuses to
start without setarch -R, so util-linux is installed and the job fails if the output contains
skipped: ThreadSanitizer.

The Rust toolchain is pinned to 1.100.0, so rustfmt and clippy cannot move the gates
under the repository without a commit.

Only clippy's correctness group is an error. The workspace carries ~30 style lints, which are
advisory; the correctness group catches a real defect rather than a preference.

The drift checks are scripts, not inline shell

python3 scripts/check_sources_listed.py   # .cpp vs SWEETLOADER_SOURCES, tests vs run_tests.sh
python3 scripts/check_pinned_deps.py      # GIT_TAG is a 40-hex commit, NOTICE matches

check_sources_listed.py catches the two failures that are otherwise silent: a new .cpp never
added to SWEETLOADER_SOURCES is simply not compiled, and a new host test never registered never
runs.

gen_rules.py had no test at all

It is what turns a Minecraft build into a hook address, and its premise is that .data.rel.ro
holds pre-relocation zeros offline, so the typeinfo and vtable pointers must be rebuilt from
.rela.dyn before anything is read. The new test builds a synthetic stripped-binary-shaped ELF
with the same section shape as the game binary and checks the derivation end to end.
Mutation-checked: skipping the relocation step fails 3 of the 8 tests.

Verification

The final commit is green on all four jobs, on GitHub's runners:

Job	Result	Notes
host-tests	pass	57 tests, three TSAN banners present, no skips
rust	pass	fmt clean, correctness clippy exit 0, 249 tests
native-build	pass	NDK r28c build; AArch64, 158 JNI exports, SDK symbols present, no archive leakage
repo-integrity	pass	both checkers and all three tooling tests

actionlint 1.7.7 with shellcheck 0.10.0 is clean on the workflow, and shellcheck is clean on
tests/cosmetics/run_tests.sh.

One residual risk worth stating: the gates assert the library exports the symbols the headers
declare, but nothing on a runner loads it into a JVM or an arm64 device. That needs a device or
an emulator, and is out of scope here.